Google's eighth-generation TPU splits into two architectures for the first time. STT breaks down how TPU 8t removes the training I/O bottleneck, and how TPU 8i uses large on-chip SRAM and OCS optical circuit switching to cut MoE inference network hops by 56%.

The OCP 2026 white paper, Lumentum's capacity lock-up race and Anthropic's 1-million-TPU deal mark the arrival of the optical direct-connect era. The full edition adds STT's golden-ratio formula, cube-position formula, a $22k OCS BOM, an 800G/1.6T BOM comparison and 13 stock view cards.

With $16.7B of AI revenue in a single quarter, Broadcom made the case that custom XPUs are a main line of their own, and broke precedent with a two-year path to $230B of FY2028 AI revenue. But the real information was in what Hock Tan volunteered: the chips can ship, but the data centers may not get built in time. Six customers, 30 GW, and the bottleneck isn't semiconductors.

OCS has gone from Google's secret weapon to the OCP community's officially endorsed next-generation data center backbone. MEMS is shipping today, silicon photonics OCS is where the decisive battle will be fought, and GPU port-level reconfiguration is the next frontier.